AI Infrastructure Runs on Different Clocks

AI infrastructure is often treated as one investment category.

We think that is increasingly a capital-allocation mistake.

A datacenter can remain useful for decades. GPUs can become economically obsolete within a few years. Models can be replaced faster still.

Those assets may sit in the same AI stack, but they should not be financed, valued or managed on the same clock.

Microsoft’s latest results make the distinction unusually clear. Of $41 billion in quarterly capital expenditure, roughly two-thirds went into short-lived assets, primarily CPUs and GPUs. The company is also designing its AI architecture so that the harness, context, memory and action space sit outside any single model family — making models substitutable.

That points to a simple principle:

The faster an asset becomes obsolete, the shorter the payback discipline should be.

Physical infrastructure can justify long-duration capital. Compute needs tighter return thresholds. Models should rarely be treated as durable strategic assets at all.

The durable value should sit around them: proprietary context, workflows, data, memory, governance and the operating capability to replace one model with another.

This matters more as AI investment moves deeper into credit markets and investors become more selective about the returns behind large infrastructure programmes.

The key question is no longer simply how much AI infrastructure should be built.

It is which layer deserves long-duration capital — and which layer is simply a rapidly depreciating production input.

In fast technology cycles, one of the most expensive mistakes is financing a temporary advantage as if it were a permanent asset.